Since 18 of December 2019 uses Nucleus credentials. Visit our help pages for information on how to Register and Sign-in using Nucleus.
22-27 October 2018
Mahatma Mandir Conference Centre
Asia/Kolkata timezone

Machine learning for disruption warning on Alcator C-Mod, DIII-D, and EAST Tokamaks

25 Oct 2018, 14:00
4h 45m
Mahatma Mandir Conference Centre

Mahatma Mandir Conference Centre

Gandhinagar (nearest Airport: Ahmedabad), India
Poster P6 Posters


Dr Robert Granetz (MIT)


We find that disruption prediction using machine learning (ML), trained on large databases containing only plasma parameters that are available in real time on C-Mod, DIII-D, and EAST, differ substantially in performance among the three machines, implying that a universal real time disruption warning algorithm may be problematic. This could have important implications for disruption prediction and avoidance on ITER, for which development of a training database of disruptions may be infeasible. Whether or not disruption prediction can be improved by incorporating additional real time measurements, or with more sophisticated AI methods, is unclear. The database for each tokamak contains parameters sampled at ~$10^6$ times throughout ~$10^4$ discharges, disruptive and non-disruptive, over the last 3-4 years of operation. We find that a number of parameters (e.g. $P_{\rm rad}$/$P_{\rm input}$, $\ell_{\rm i}$, $n$/$n_{\rm G}$, $B_{n=1}$/$B_{\rm T}$) exhibit changes as a disruption is approached on one or more of these tokamaks. However, the details of these precursor behaviors are markedly different on each machine. We use a shallow ML method known as Random Forests, applied to a binary classification scheme. We define the two classes as "*close to a disruption*” and "*far from a disruption or from a non-disruptive shot*”. The threshold time that divides "close” from "far” is determined by optimising the classification prediction accuracy for each machine. We find that the timescales of disruption warning behavior are very different for the different machines, and that the fraction of correctly predicted disruption samples varies considerably, ranging from 74% for DIII-D, to just 35% for C-Mod. For C-Mod in particular, it is difficult to predict upcoming disruptions more than just a few milliseconds in advance. This work supported in part by: US DoE Grants DE-FC02-99ER54512, DE-SC0010720, DE-SC0010492, DE-FC02-04ER54698, DE-SC0014264 National Natural Science Foundation of China Grants 11475002, 11775262, 11475224, 11575247, 11475225, 11775266 and 11505235 National Magnetic Confinement Fusion Science Program of China Grants 2014GB103000 and 2015GB102004
Paper Number EX/P6-20
Country or International Organization United States of America

Primary author


Prof. BiagJia Xiao (CnIPPCAS) Dr Biao Shen (CnIPPCAS) Dr Cristina Rea (MIT) Dr DaLong Chen (CnIPPCAS) Mr Kevin Montes (MIT) Dr Nicholas Eidietis (General Atomics) Dr Orso Meneghini (General Atomics)

Presentation Materials